Microsoft’s latest medical AI system diagnosed more than eight in 10 complex cases in controlled tests, substantially outperforming generalist physicians working through the same cases. But whether it can deliver similar results in hospitals remains an open question.
The system, called Microsoft AI Diagnostic Orchestrator (MAI-DxO), correctly diagnosed 83.2% of complex cases drawn from The New England Journal of Medicine, compared with 36.1% for generalist physicians who could use online search and reference materials.
Physicians working without those resources scored 22.3%, according to research published in Science on Oct. 22, 2026.
For healthcare providers considering AI-assisted diagnosis, the results offer a promising glimpse of the technology’s potential. However, the research relied on historical medical cases rather than actual patient encounters, leaving important questions about clinical reliability, safety, and implementation.
Microsoft uses AI agents to simulate diagnostic reasoning
MAI-DxO takes a unique approach from AI tools that generate a diagnosis after receiving a complete set of symptoms and test results. It coordinates multiple AI agents to work through a case in stages, determining which questions to ask, selecting tests, interpreting results, and updating its possible diagnoses as more information becomes available.
According to Microsoft, the system uses OpenAI’s GPT-5 for the primary evaluation. The orchestrator coordinates the diagnostic process, simulating how a panel of clinical experts might work through a complicated case.
Researchers evaluated MAI-DxO using the Sequential Diagnosis Benchmark, or SDBench. The assessment included 304 complex cases drawn from clinical discussions published in The New England Journal of Medicine and 325 emergency department cases from the MIMIC-IV-ED database.
On the emergency department cases, MAI-DxO achieved 80.3% diagnostic accuracy, compared with 46.8% for physicians. The evaluation also considered the estimated costs of the diagnostic processes, measuring the resources associated with the tests and procedures selected during each workup.
Benchmark results do not establish clinical readiness
The accuracy figures may be promising, but that still doesn’t mean Microsoft‘s system is better than doctors in every clinical scenario.
The study compared the system with individual generalist physicians on historical cases presented through a controlled, text-only evaluation. It did not establish that MAI-DxO outperforms medical specialists, multidisciplinary clinical teams, or physicians treating patients in real-world settings.
Actual patient encounters introduce challenges that written benchmarks cannot fully reproduce. Patients may provide incomplete medical histories, symptoms can change, and physicians must consider information beyond documented test results.
The study also measured diagnostic accuracy rather than treatment decisions, patient safety, or clinical outcomes.
Microsoft is now looking to evaluate the technology in a clinical setting. Reuters reported that the company plans to work with Beth Israel Deaconess Medical Center in Boston to assess the system’s ability to triage patients in emergency department waiting rooms.
That evaluation could provide more evidence about the technology’s practical usefulness, although successful triage would not, by itself, establish readiness for broader diagnostic deployment.
Estimated cost advantages need further testing
Microsoft also reported that MAI-DxO could achieve strong diagnostic accuracy with lower estimated testing costs in some configurations. In one budget-constrained evaluation involving complex cases, the system achieved 80.9% accuracy at an estimated cost of $2,241 per case. The physicians’ average estimated cost was $4,725.
These findings could be relevant to healthcare organizations assessing AI tools for clinical decision support. However, the study’s cost figures are estimates based on diagnostic testing, not measurements of actual hospital spending, implementation costs, or patient outcomes.
The results therefore do not establish that deploying MAI-DxO would reduce overall healthcare expenditure. Its practical financial impact would depend on how the system performs in clinical settings and how it fits into existing diagnostic processes.
More Microsoft news
- Inside Microsoft’s Real-Time War Against Cybersecurity Threats
- Project Ire: Microsoft Tests AI That Autonomously Detects Malware
- Microsoft Targets ‘Critical AI Talent’ from Meta to Dominate Next AI Breakthroughs
- Windows 10 Support Ends Soon, Though Extended Security Updates Offers Are Available
Microsoft targets broader medical AI capabilities
Microsoft is presenting MAI-DxO as part of a bigger effort to develop what AI chief Mustafa Suleyman calls “medical superintelligence.” The company’s longer-term vision extends beyond diagnosis to other healthcare activities, including treatment decisions and clinical care coordination.
Those capabilities represent Microsoft’s ambitions for future AI development, rather than functions established by the MAI-DxO benchmark. The current study focuses on diagnostic reasoning, and further research will be needed to determine how the system performs in practical healthcare applications.
For healthcare technology leaders, the top question is how well AI diagnostic systems can demonstrate consistent benefits outside controlled evaluations.
Microsoft’s findings suggest that an orchestrated AI system can solve many of the cases it was evaluated on, but they don’t yet show whether using the system leads to better care.
The next test for Microsoft’s system is not simply whether it can diagnose difficult cases correctly, but whether it can help clinicians make safer, more reliable decisions in actual healthcare settings.
Other news: A new study examines how continuous glucose monitors could help people with Type 2 diabetes better track their blood sugar levels and manage their condition.